Defects in the form of structural flaws, process residues, and external contamination occur during the production of semiconductor wafers. Defects are typically detected by a class of instruments called defect scanners. Such instruments automatically scan wafer surfaces and detect optical anomalies using a variety of techniques. The location of these anomalies with respect to the pattern of semiconductor devices on the wafer surface is recorded. This information, or “defect map,” is stored in a computer file and sent to a defect review station.
Using the defect map to locate each defect, a human operator observes each defect under a microscope and characterizes each defect according to type (e.g., particle, pit, scratch, or contaminant). Information gained from this process is used to correct the source of defects, and thereby improve the efficiency and yield of the semiconductor production process. Unfortunately, people are relatively slow and are quickly fatigued by the highly-repetitive task of observing and characterizing defects.
Methods of automatically characterizing defects, collectively known as Automatic Defect Characterization, or “ADC,” have been developed to overcome the disadvantages of manual defect characterization. Conventional white-light-microscope-based review stations are automated to load a wafer that has been mapped for defect location by a defect scanner. Once the mapped wafer is loaded, the review station:                1. positions the wafer to image the site of a defect, as indicated by the defect map;        2. focuses on the site of the defect;        3. captures a digital image of the site using a digital TV camera;        4. processes and analyzes the captured image of the site to locate the defect; and        5. further analyzes the data to characterize the defect.The above process is repeated for each defect (or a predetermined subset of defects) on the wafer. The wafer is then unloaded and the process is repeated for another wafer. By eliminating a fatiguing and highly repetitive task, such automated review stations reduce labor costs and provide improved consistency and accuracy over human operators.        
Conventional ADC systems capture a conventional white-light microscope image as an array A representing a two-dimensional image. The image is an x-y array of n by m pixels, where typical values might be n=640, m=480, or n=512, m=512. This array may be represented as:                A(x, y, Ir, Ig, Ib),where x and y are pixel coordinates, and Ir, Ig, and Ib represent the intensities of the red, green, and blue image components, respectively. Of course, grey scale images may also be used, as may other color schemes, such as those of the YUV and YIQ commercial standard formats. In the case of a gray scale image, a single intensity parameter Ig is used.        
In addition to imaging the defect site, at least one reference image Aref is also stored. The reference image may be a previously stored data-base image of a known-good area of the same or a similar die on the same or on a similar wafer, or it may be a specific image taken from, e.g., an adjacent die. The reference image is compared with the image containing the defect. Any differences measured between the two images will indicate the location and extent of the defect.
Multiple reference images are usually required because slight differences in focus position between the reference and test images may cause false discrepancies to appear. In some cases, a separate reference image is not taken, and instead the reference image is a portion of the same image containing the defect, but from a region of the image where no defect occurs. In general, this latter method is faster but less reliable than methods that use a separate reference image, and works only for images containing repetitive structures or patterns.
Several conventional techniques are available to process images for automatic defect characterization. One such technique is described by Youling Lin, M. S., in Techniques for Syntactic Analysis of Images with Application for Automatic Visual Inspection, a dissertation in business administration submitted in December of 1990 to the graduate faculty of Texas Tech University in partial fulfillment of the requirements of the degree of doctor of philosophy, which is incorporated herein by this reference.
Lin describes ADC techniques for processing a two-dimensional microscope image. According to Lin, low-level image processing enhances surface features and reduces noise. This process is performed on intensity (gray scale) variations of the image. Lin describes an extreme-median digital filter to accomplish this task.
Next, Lin describes techniques for identifying feature boundaries and converting the boundaries into a list of symbolic geometric “primitives.” Suppose, for example, that a surface feature has the shape of a half-circle. Such a feature will have a boundary shaped approximately like the letter “D.” This boundary could be converted into two geometric primitives; a line segment (specified by length and direction) representing the vertical portion of the “D,” and an arc (specified by position and radius) representing the curved portion of the letter “D.” More complex shapes may be similarly represented using a large number of connected line segments, angles, and arcs.
Symbolic geometric primitive extraction is performed, for example, by statistical comparison of the edge data with a representation of geometric primitives, such as line segments, arcs, or angles. The surface-feature boundary data is replaced with a set of primitives that best describes the boundary.
The preceding steps are performed both for at least one reference image and for a test image. Then, using techniques derived from compiler theory, the set of reference primitives is compared, primitive by primitive, with the set of test primitives. When a discrepancy is encountered between the sets of reference and test primitives, a rule-based expert system notes the discrepancy and continues the comparison. The discrepancies (i.e., the differences between the sets of reference and test primitives) define the location of a defect.
Alternatively, the defect area may be located by overlaying the test and reference images, aligning them by correlation techniques, and subtracting the images one from the other. Defects will show up as areas where the test and reference images have large difference values.
Having identified the location of a defect, the boundaries of the defect are identified and represented by a set of primitives in the manner described above for the test and reference images. In one embodiment, where more than one defect is located in a single image, only the defect with the largest area is selected for further processing.
Next, the set of primitives representing the image portion containing the defect is used to develop a set of defect parameters, each defect parameter representing a single feature of the defect. For example, one defect parameter may represent the area of the defect and another the shape of the defect. Moreover, characteristics of the area defined by the defect boundaries may be used to derive additional defect parameters. For example, the defect area may be analyzed for average intensity, variations in intensity from one pixel to the next or within a small region (“texture”), color, or color coordinates. The defect parameters are conventionally expressed in a normalized form so that they run from, e.g., 0 to 1 or −1 to 1. A defect-parameter vector is then defined by these parameters.
The defect-parameter vector is compared, using conventional fuzzy logic techniques, with typical vectors for each known type of defect. Based on this comparison, the ADC system characterizes the defect and estimates the probability that the selected characterization is accurate. For a more detailed description of one method of developing a defect-parameter vector, see “Techniques for Syntactic Analysis of Images with Application for Automatic Visual Inspection,” which is incorporated herein by reference.
For further discussion of conventional ADC techniques, see the IBM technical disclosure entitled “Automated Classification of Defects in Integrated Circuit Manufacturing,” by Frederick Y. Wu, et al., which is incorporated herein by this reference.
Conventional ADC images have a number of shortcomings. For example, small pits versus particles cannot be distinguished, shallow structures are not discernible, and subsurface defects cannot be characterized. And, if a defect or structure on a surface is “tall,” focusing on one level leaves other levels out of focus. Accuracy of the automatic focus between the test and reference image then becomes critical because small variations in focus cause the boundary between two structures of different heights to change in appearance. A conventional ADC system may then interpret this variation as a potential defect when it is not. Human operators can compensate for this to some degree by, e.g., moving the focus up and down and interpreting three-dimensional aspects of the images, but this wastes valuable time. Moreover, if there are low optical contrasts between the defect and the surrounding material (e.g., the defect is of approximately the same color or reflective intensity as the surrounding surface of the semiconductor), an ADC scheme can fail to detect the true shape—or even the existence—of the defect. Therefore, what is needed is a more accurate method of automatically characterizing defects.